apply suggestions from gemini

Author: chenyushuo

trinity/common/config.py

@@ -931,6 +931,8 @@ def _check_explorer_input(self) -> None:
             dataset.batch_size = self.buffer.batch_size
             if not dataset.name:
                 dataset.name = f"eval_taskset_{idx}"
+            if len(dataset.eval_at_k) == 0:
+                dataset.eval_at_k = [1]
             dataset.repeat_times = max(dataset.eval_at_k)
 
             # eval_workflow has higher priority than workflow in eval tasksets, so we set it first

trinity/explorer/scheduler.py

@@ -58,8 +58,8 @@ def calculate_task_level_metrics(
                 if k > len(values):
                     continue
                 result[f"{key}/mean@{k}"] = sum(values[:k]) / k
-                result[f"{key}/best@{k}"] = max(values[:k]) / k
-                result[f"{key}/worst@{k}"] = min(values[:k]) / k
+                result[f"{key}/best@{k}"] = max(values[:k])
+                result[f"{key}/worst@{k}"] = min(values[:k])
         return result
     else:
         return {